Merge pull request #21 from fuyw/add_model_name_in_compilation

Add ability to specify root function's name in compiled binary

lleaves/compiler/codegen/codegen.py

@@ -39,7 +39,7 @@ class LTree:
     class_id: int
 
 
-def gen_forest(forest, module, fblocksize):
+def gen_forest(forest, module, fblocksize, froot_func_name):
     """
     Populate the passed IR module with code for the forest.
 
@@ -81,7 +81,7 @@ def gen_forest(forest, module, fblocksize):
     root_func = ir.Function(
         module,
         ir.FunctionType(ir.VoidType(), (DOUBLE_PTR, DOUBLE_PTR, INT, INT)),
-        name="forest_root",
+        name=froot_func_name,
     )
 
     def make_tree(tree):

lleaves/compiler/tree_compiler.py

@@ -7,12 +7,18 @@
 from lleaves.compiler.codegen import gen_forest
 
 
-def compile_to_module(file_path, fblocksize=34, finline=True, raw_score=False):
+def compile_to_module(
+    file_path,
+    fblocksize=34,
+    finline=True,
+    raw_score=False,
+    froot_func_name="forest_root",
+):
     forest = parse_to_ast(file_path)
     forest.raw_score = raw_score
 
     ir = llvmlite.ir.Module(name="forest")
-    gen_forest(forest, ir, fblocksize)
+    gen_forest(forest, ir, fblocksize, froot_func_name)
 
     ir.triple = llvm.get_process_triple()
     module = llvm.parse_assembly(str(ir))

lleaves/lleaves.py

@@ -85,6 +85,7 @@ def compile(
         fblocksize=34,
         fcodemodel="large",
         finline=True,
+        froot_func_name="forest_root",
     ):
         """
         Generate the LLVM IR for this model and compile it to ASM.
@@ -107,6 +108,8 @@ def compile(
             very large forests.
         :param finline: Whether or not to inline function. Setting this to False will speed-up compilation time
             significantly but will slow down prediction.
+        :param froot_func_name: Name of entry point function in the compiled binary. This is the function to link when
+            writing a C function wrapper. Defaults to "forest_root".
         """
         assert 0 < fblocksize
         assert fcodemodel in ("small", "large")
@@ -117,6 +120,7 @@ def compile(
                 raw_score=raw_score,
                 fblocksize=fblocksize,
                 finline=finline,
+                froot_func_name=froot_func_name,
             )
         else:
             # when loading binary from cache we use a dummy empty module
@@ -128,7 +132,7 @@ def compile(
         )
 
         # Drops GIL during call, re-acquires it after
-        addr = self._execution_engine.get_function_address("forest_root")
+        addr = self._execution_engine.get_function_address(froot_func_name)
         self._c_entry_func = ENTRY_FUNC_TYPE(addr)
 
         self.is_compiled = True

tests/test_compile_flags.py

@@ -74,3 +74,19 @@ def test_no_inline(NYC_data):
         llvm_model.predict(NYC_data[:1000], n_jobs=2),
         lgbm_model.predict(NYC_data[:1000], n_jobs=2),
     )
+
+
+def test_function_name():
+    llvm_model = Model(model_file="tests/models/tiniest_single_tree/model.txt")
+    lgbm_model = Booster(model_file="tests/models/tiniest_single_tree/model.txt")
+    llvm_model.compile(froot_func_name="tiniest_single_tree_123132_XXX-")
+
+    data = [
+        [1.0] * 3,
+        [0.0] * 3,
+        [-1.0] * 3,
+    ]
+    np.testing.assert_almost_equal(
+        llvm_model.predict(data, n_jobs=2),
+        lgbm_model.predict(data, n_jobs=2),
+    )